一种基于预取的集群服务器调度算法

A Scheduling Algorithm Based on Web Prefetching for Cluster Servers

针对集群服务器LARD调度算法只能利用已有缓存的问题,提出一种基于预取的算法Prefetch-LARD.该算法从Web访问日志中挖掘页面之间的转移概率,建立马尔科夫链模型,在调度请求时利用概率关系提前将下一次可能访问的文档从节点磁盘取到本地cache中,提高了请求的缓存命中率;算法还采用了加权的节点超载判断方法,以提高集群节点的负载均衡度.实验表明,在同样的测试环境下,Prefetch-LARD算法比LARD算法的缓存命中率提高26.9%,系统的吞吐量相应提高18.8%.

To the problem that the scheduling algorithm of locality-aware request distribution （LARD） for cluster server can only make use of the existing node caches, an advanced algorithm based on Web prefetching, Prefetch- LARD, is proposed. By mining the transition probability between pages from Web access logs, the algorithm builds up a prefetching model based on Markov chain to fetch documents for next possible requests ahead from disks to caches. Furthermore, the algorithm adopts a weighted node choosing method to improve the load balancing metric among nodes. Experiments show that, Prefetch-LARD algorithm increases cache hit ratio up to 26.9% and the throughput up to 18.8% compared with LARD algorithm.